Parallel

API Documentation: Parallel

When browsing the various Monads included in Cats, you may have noticed that some of them have data types that are actually of the same structure, but instead have instances of Applicative. E.g. Either and Validated.

This is because defining a Monad instance for data types like Validated would be inconsistent with its error-accumulating behaviour. In short, Monads describe dependent computations and Applicatives describe independent computations.

Sometimes however, we want to use both in conjunction with each other. In the example of Either and Validated it is trivial albeit cumbersome to convert between the two types. Below is a short example of a situation where we might run into this. For simplicity, we'll use String as our type to represent errors.

import cats.syntax.all._
import cats.data._

case class Name(value: String)
case class Age(value: Int)
case class Person(name: Name, age: Age)

def parse(s: String): Either[NonEmptyList[String], Int] = {
  if (s.matches("-?[0-9]+")) Right(s.toInt)
  else Left(NonEmptyList.one(s"$s is not a valid integer."))
}

def validateAge(a: Int): Either[NonEmptyList[String], Age] = {
  if (a > 18) Right(Age(a))
  else Left(NonEmptyList.one(s"$a is not old enough"))
}

def validateName(n: String): Either[NonEmptyList[String], Name] = {
  if (n.length >= 8) Right(Name(n))
  else Left(NonEmptyList.one(s"$n Does not have enough characters"))
}

Now we want to parse two Strings into a value of Person:

def parsePerson(ageString: String, nameString: String) =
  for {
    age <- parse(ageString)
    person <- (validateName(nameString).toValidated, validateAge(age).toValidated)
       .mapN(Person)
       .toEither
  } yield person

We had to convert to and from Validated manually. While this is still manageble, it gets worse the more Eithers we want to combine in parallel.

To mitigate this pain, Cats introduces a type class Parallel that abstracts over Monads which also support parallel composition. It is simply defined in terms of conversion functions between the two data types:

trait Parallel[M[_]] {
  type F[_]

  def sequential: F ~> M

  def parallel: M ~> F
}

Where M[_] has to have an instance of Monad and F[_] an instance of Applicative.

Recall that ~> is just an alias for FunctionK. This allows us to get rid of most of our boilerplate from earlier:

def parsePerson(ageString: String, nameString: String) =
  for {
    age <- parse(ageString)
    person <- (validateName(nameString), validateAge(age)).parMapN(Person)
  } yield person

We can also traverse over a Traverse using Parallel:

List(Either.right(42), Either.left(NonEmptyList.one("Error 1")), Either.left(NonEmptyList.one("Error 2"))).parSequence
// res0: Either[NonEmptyList[String], List[Int]] = Left(
//   value = NonEmptyList(head = "Error 1", tail = List("Error 2"))
// )

Parallel is also really useful for zipping collections. The standard Applicative instance for List, Vector, etc. behaves like the cartesian product of the individual collections:

(List(1, 2, 3), List(4, 5, 6)).mapN(_ + _)
// res1: List[Int] = List(5, 6, 7, 6, 7, 8, 7, 8, 9)

However often we will want to zip two or more collections together. We can define a different ap for most of them and use the parMapN syntax for that:

(List(1, 2, 3), List(4, 5, 6)).parMapN(_ + _)
// res2: List[Int] = List(5, 7, 9)

NonEmptyParallel - a weakened Parallel

Some types cannot form a Monad or an Applicative because it's not possible to implement the pure function for them. However, these types can often have instances for FlatMap or Apply. For types like these we can use the NonEmptyParallel type class. An example for one of these is ZipList.

With instances of NonEmptyParallel it's not possible to use the parTraverse and parSequence functions, but we can still use parMapN and also parNonEmptyTraverse and parNonEmptySequence, which are analogous to the functions defined on NonEmptyTraverse.